Claude artifacts: what they are, and how to check one before you share it

An operations manager pastes three months of supplier delivery logs into Claude and asks for a one-page summary for Monday's meeting. The analysis scrolls past in the chat. The summary doesn't: it opens in its own panel, formatted, editable, with an export button. That panel is an artifact.

Anthropic's own definition is the clearest one: an artifact is "anything Claude makes for you that you'd put in front of someone" (Anthropic Help Centre, What are artifacts). The chat is the working; the artifact is the deliverable. You can come back to it, edit it by hand or by asking Claude, and share it.

Artifacts first appeared as a preview alongside Claude 3.5 Sonnet in June 2024 (Anthropic, June 2024). They have changed a lot since. Today they come in five types:

Small interactive tools (a pricing calculator, a sortable table) also count. Connecting an artifact to your apps and having it store data needs Pro, Max, Team or Enterprise.

One recent change catches people out. Artifacts made in chat before 16 September 2026 are now "legacy artifacts": they still work and can still be shared, but you can't create new ones in that format (Anthropic Help Centre, checked October 2026).

Most people get their first artifact by accident. To get one on purpose:

1. Switch on file creation. Artifacts depend on "Cloud code execution and file creation". On Free, Pro and Max it lives in Settings > Capabilities; on Team and Enterprise an owner controls it in Organization settings > Capabilities.
2. Ask for the thing, not the text. "Turn this into a two-slide summary for the leadership meeting" gets you a deck. "Summarise this" gets you a paragraph in the chat. Alternatively, start from a template in the Artifacts tab (Docs, Slides, Design, Dashboards, Motion).
3. Edit in place. Type into the artifact directly, ask Claude for changes in the chat, or highlight a passage and choose "Edit with Claude" to rework only that part.
4. Export or share. Download in the formats above, or share the link. Artifacts start private, so nobody sees one until you choose.

A prompt that gets a useful first draft (illustrative):

Using the delivery log above, make a one-page doc for the ops leadership meeting.
Lead with the three suppliers with the most late deliveries, by count and by
share of their deliveries. Add a short table. Under the table, list every
assumption you made about the data (date ranges, how "late" is defined,
rows you excluded).

That last instruction is the one most people skip. It costs a sentence, and it puts the decisions Claude made into the document itself, where your reader can see them.

An artifact shows what Claude decided. It doesn't show what each number rests on. Count does that part: connected to Claude, it tags every claim in the answer as a citation, an assumption, model knowledge or a hallucination.

We like artifacts. We've argued that AI analysis tools should produce artifacts, not transcripts, because a document you can return to, edit and hand over is worth more than a long scroll of chat. Claude's artifacts do that well.

Where they shine

• Turning finished thinking into a format. Once you know what the memo says, getting it into a clean Word doc or a six-slide deck is minutes, not an afternoon.
• Iterating on structure. "Move the risks section up, cut the appendix, turn the table into a chart" is the kind of instruction Claude follows reliably.
• Small tools. A margin calculator for the sales team or a filterable list of open contracts is a reasonable afternoon's artifact, and nobody has to file a ticket.

Where they make things worse

Polish. A number in a chat reply looks provisional; the same number on a formatted slide looks approved. The artifact separates the conclusion from the working that produced it, and the reader of a deck rarely scrolls back through the chat. That is fine when the working was right, and quietly expensive when it wasn't.

The second problem is drift. Ask for ten rounds of edits and you will usually get them, but each round is another chance for a figure to change without anyone asking it to. Re-check numbers after heavy editing, not only after the first draft.

The limits are modest, and mostly matter once an artifact stores data or calls Claude itself (Anthropic Help Centre, checked October 2026):

Two practical points follow. First, the first time an artifact uses shared storage you'll see a confirmation; Anthropic's advice is to "check whether the artifact uses shared storage" before entering anything sensitive. A team survey tool that writes to shared storage is, by design, readable by the whole team.

Second, be careful with other people's published artifacts. Building on one brings their code into your chat, and Anthropic says to do that only with artifacts from people you trust. Each person connects their own apps, so sharing an artifact doesn't share your Google Drive access, which is the right default.

Anthropic doesn't publish accuracy figures for artifacts, and its help pages are frank about the gaps. When an artifact throws an error, "Try fixing with Claude" is offered with the caveat that "Claude will attempt to fix the error, but success isn't guaranteed". On file creation more broadly: "You may need to review and refine Claude's outputs to meet your exact requirements."

Independent tests explain the caution. On SpreadsheetBench V2, which asks AI systems to complete 321 end-to-end business workflows such as financial models, debugging and visualisation, the best published score is 34.89% (SpreadsheetBench, checked October 2026). A dashboard artifact is exactly that kind of workflow. The chart will render either way.

A two-minute check before you share

1. Pick the headline number (the one in the title or the first slide) and ask: "Show me the calculation behind this figure, with the rows it used." If it can't, treat the number as unverified.
2. Recompute one figure yourself from the source data. One, not all. If one is wrong, the others need looking at.
3. Check the edges of every chart: the date range, the filters, and whether the segments add up to the total shown.
4. Ask Claude: "Which numbers or statements in this artifact did not come from the data I gave you?" Anything it names is either model knowledge or an assumption.
5. Diff after edits. If you asked for several rounds of changes, compare the key figures in the final version against the first draft.

Steps 1 and 4 are what a reasoning chain does for every claim at once. Count connects to Claude, splits an answer into claims, and labels each one as a citation, an assumption, model knowledge or a hallucination, so the slide and the working behind it travel together.

Our working pattern for anything that leaves the team: let Claude draft the artifact, run the check above, and only then share the link. The artifact is excellent at the last mile, the formatting, the restructuring, the export to PowerPoint. It is not where the analysis should be judged, because by the time something is on a slide it has lost the context that would tell you whether to believe it.

If the artifact is built on your company's data, the steps that matter happen earlier, in Claude data analysis or through connectors. If it is built on web research, see Claude Research for how to check the citations it rests on. People tend to check fluent, finished work least. An artifact is about as fluent and finished as AI output gets.

What are artifacts in Claude?

Are Claude artifacts free?

How do I turn on artifacts in Claude?

Who can see my Claude artifacts?